Content-based Image Retrieval (CBIR) · 2006. 4. 17. · Content-based Image Retrieval (CBIR)...

1

Content-based Image Retrieval (CBIR)

Searching a large database for images that match a query:

What kinds of databases?What kinds of queries?What constitutes a match?How do we make such searches efficient?

2

Applications

Art Collections e.g. Fine Arts Museum of San FranciscoMedical Image DatabasesCT, MRI, Ultrasound, The Visible HumanScientific Databasese.g. Earth SciencesGeneral Image Collections for LicensingCorbis, Getty ImagesThe World Wide Web

3

What is a query?

an image you already have

a rough sketch you draw

a symbolic description of what you wante.g. an image of a man and a woman ona beach

SYSTEMS

4

5

Some Systems You Can Try

Corbis Stock Photography and Pictures

http://pro.corbis.com/

• Corbis sells high-quality images for use in advertising,marketing, illustrating, etc.

• Search is entirely by keywords.

• Human indexers look at each new image and enter keywords.

• A thesaurus constructed from user queries is used.

6

QBICIBM’s QBIC (Query by Image Content)

http://wwwqbic.almaden.ibm.com

• The first commercial system.

• Uses or has-used color percentages, color layout,texture, shape, location, and keywords.

7

BlobworldUC Berkeley’s Blobworld

http://elib.cs.berkeley.edu/blobworld

• Images are segmented on color plus texture

• User selects a region of the query image

• System returns images with similar regions

• Works really well for tigers and zebras

8

DittoDitto: See the Web

http://www.ditto.com

• Small company

• Allows you to search for pictures from web pages

9

Image Features / Distance Measures

Query Image Retrieved Images

Image Database

Distance Measure

Image Feature

Extraction

Feature SpaceImages

User

10

Features

• Color (histograms, gridded layout, wavelets)

• Texture (Laws, Gabor filters, local binary pattern)

• Shape (first segment the image, then use statisticalor structural shape similarity measures)

• Objects and their Relationships

This is the most powerful, but you have to be able torecognize the objects!

11

Color Histograms

12

QBIC’s Histogram Similarity

The QBIC color histogram distance is:

dhist(I,Q) = (h(I) - h(Q)) A (h(I) - h(Q))T

• h(I) is a K-bin histogram of a database image

• h(Q) is a K-bin histogram of the query image

• A is a K x K similarity matrix

13

Similarity Matrix

R G B Y C V1 0 0 .5 0 .50 1 0 .5 .5 00 0 1

11

1

RGBYCV

How similar is blue to cyan?

??

14

Gridded Color

Gridded color distance is the sum of the color distancesin each of the corresponding grid squares.

1 2

3 34 4

1 2

What color distance would you use for a pair of grid squares?

15

Color Layout (IBM’s Gridded Color)

16

Texture Distances

• Pick and Click (user clicks on a pixel and systemretrieves images that have in them a region withsimilar texture to the region surrounding it.

• Gridded (just like gridded color, but use texture).

• Histogram-based (e.g. compare the LBP histograms).

17

Laws Texture

18

Shape Distances

• Shape goes one step further than color and texture.

• It requires identification of regions to compare.

• There have been many shape similarity measuressuggested for pattern recognition that can be usedto construct shape distance measures.

19

Global Shape Properties:Projection Matching

041320

Feature Vector(0,4,1,3,2,0,0,4,3,2,1,0)

0 4 3 2 1 0

In projection matching, the horizontal and verticalprojections form a histogram.

What are the weaknesses of this method? strengths?

20

Global Shape Properties:Tangent-Angle Histograms

135

0 30 45 135

Is this feature invariant to starting point?Is it invariant to size, translation, rotation?

21

Boundary Matching• Fourier Descriptors

• Sides and Angles

• Elastic Matching

The distance between query shape and image shapehas two components:

1. energy required to deform the query shape intoone that best matches the image shape

2. a measure of how well the deformed query matchesthe image

22

Del Bimbo Elastic Shape Matching

query retrieved images

23

Regions and Relationships

• Segment the image into regions

• Find their properties and interrelationships

• Construct a graph representation withnodes for regions and edges for spatial relationships

• Use graph matching to compare images

Like what?

24

Tiger Image as a Graph

sky

sand

tiger grass

aboveadjacent

above

inside

above aboveadjacent

image

abstract regions

25

Object Detection:Rowley’s Face Finder

1. convert to gray scale2. normalize for lighting*

3. histogram equalization4. apply neural net(s)

trained on 16K images

What data is fed tothe classifier?

32 x 32 windows ina pyramid structure

* Like first step in Laws algorithm, p. 220

26

Fleck and Forsyth’s Flesh DetectorThe “Finding Naked People” Paper

• Convert RGB to HSI• Use the intensity component to compute a texture map

texture = med2 ( | I - med1(I) | )• If a pixel falls into either of the following ranges,

it’s a potential skin pixel

texture < 5, 110 < hue < 150, 20 < saturation < 60texture < 5, 130 < hue < 170, 30 < saturation < 130

median filters of radii 4 and 6

Look for LARGE areas that satisfy this to identify pornography.

27

Wavelet Approach

Idea: use a wavelet decomposition to representimages

What are wavelets?

• compression scheme

• uses a set of 2D basis functions

• representation is a set of coefficients, one foreach basis function

28

Jacobs, Finkelstein, Salesin Methodfor Image Retrieval (1995)

1. Use YIQ color space

2. Use Haar wavelets

3. 128 x 128 images yield 16,384 coefficients x 3 color channels

4. Truncate by keeping the 40-60 largest coefficients(make the rest 0)

5. Quantize to 2 values (+1 for positive, -1 for negative)

29

JFS Distance Metric

d(I,Q) = w00 | Q[0,0] - I[0,0] | + ∑ wij | Q’[i,j] - I’[i,j] |ij

where the w’s are weights,

Q[0,0] and I[0,0] are scaling coefficients relatedto average image intensity,

Q’[i,j] and I’[i,j] are the truncated, quantized coefficients.

30

Experiments

20,558 image database of paintings

20 coefficients used

User “paints” a rough version of the paintinghe /she wants on the screen.

See Video

31

Relevance Feedback

In real interactive CBIR systems, the user shouldbe allowed to interact with the system to “refine”the results of a query until he/she is satisfied.

Relevance feedback work has been done by a number of research groups, e.g.

• The Photobook Project (Media Lab, MIT)• The Leiden Portrait Retrieval Project• The MARS Project (Tom Huang’s group at Illinois)

32

Information Retrieval Model*

An IR model consists of:a document modela query modela model for computing similarity between documents and the queries

Term (keyword) weighting

Relevance Feedback

*from Rui, Huang, and Mehrotra’s work

33

Term weighting

Term weightassigning different weights for different keyword(terms) according their relative importance to the document

define to be the weight for term ,k=1,2,…,N, in the document idocument i can be represented as a weight vector in the term space

ikwkt

[ ]iNiii wwwD ;...;; 21=

34

Term weighting

The query Q also is a weight vector in the term space

The similarity between D and Q

[ ]qNqq wwwQ ;...;; 21=

QDQDQDSim

),( =.

35

Using Relevance Feedback

The CBIR system should automatically adjust the weight that were given by the user for the relevance of previously retrieved documents

Most systems use a statistical method for adjusting the weights.

36

The Idea of Gaussian Normalization

If all the relevant images have similar values for component j

the component j is relevant to the query

If all the relevant images have very different values for component j

the component j is not relevant to the query

the inverse of the standard deviation of the related image sequence is a good measure of the weight for component jthe smaller the variance, the larger the weight

37

The Leiden Portrait System was an example of use of relevance feedback.

• The user was presented with a set ofportraits on the screen

• Each portrait had a “yes” and “no” boxunder it, initialized to all “yes”

• The user would click “no” on the onesthat were not the sort of portrait desired

• The system would repeat its search withthe new feedback (multiple times if desired)

38

Mockup of the Leiden System

39

Andy Berman’s FIDS Systemmultiple distance measuresBoolean and linear combinationsefficient indexing using images as keys

40

Andy Berman’s FIDS System:

Use of key images and the triangle inequalityfor efficient retrieval.

41

Andy Berman’s FIDS System:

Bare-Bones Triangle Inequality Algorithm

Offline

1. Choose a small set of key images

2. Store distances from database images to keys

Online (given query Q)

1. Compute the distance from Q to each key

2. Obtain lower bounds on distances to database images

3. Threshold or return all images in order of lower bounds

42

Andy Berman’s FIDS System:

43

Andy Berman’s FIDS System:

Bare-Bones Algorithm with Multiple Distance Measures

Offline

1. Choose key images for each measure

2. Store distances from database images to keys for all measures

Online (given query Q)

1. Calculate lower bounds for each measure

2. Combine to form lower bounds for composite measures

3. Continue as in single measure algorithm

44

Andy Berman’s FIDS System:

Triangle Tries

A triangle trie is a tree structure that stores the distances from database images to each of the keys, one key per tree level.

root

Distance to key 143

1 9 Distance to key 28

YXW,Z

45

Andy Berman’s FIDS System:

Triangle Tries and Two-Stage Pruning

• First Stage: Use a short triangle trie.

• Second Stage: Bare-bones algorithm on the imagesreturned from the triangle-trie stage.

The quality of the output is the same as with the bare-bones algorithm itself, but execution is faster.

46

Andy Berman’s FIDS System:

Andy Berman’s FIDS System:

47

48

Andy Berman’s FIDS System:

Performance on a Pentium Pro 200-mHz

Step 1. Extract features from query image. (.02s ≤ t ≤ .25s)

Step 2. Calculate distance from query to key images.(1µs ≤ t ≤ .8ms)

Step 3. Calculate lower bound distances.(t ≈ 4ms per 1000 images using 35 keys, which is about 250,000 images per second.)

Step 4. Return the images with smallest lower bound distances.

49

Demo of FIDS

http://www.cs.washington/research/imagedatabase/demo

50

Weakness of Low-level Features

Can’t capture the high-level concepts

51

Current Research ObjectiveQuery Image Retrieved Images

Image Database

ImagesObject-oriented

FeatureExtraction

…

Animals

Buildings

Office Buildings

Houses

Transportation

•Boats

•Vehicles

…

boat

Categories

User

52

Overall Approach

• Develop object recognizers for common objects

• Use these recognizers to design a new set of both

low- and mid-level features

• Design a learning system that can use these

features to recognize classes of objects

53

Boat Recognition

54

Vehicle Recognition

55

Building Recognition

56

Building Features: Consistent Line Clusters (CLC)A Consistent Line Cluster is a set of lines that are homogeneous in terms of some line features.

Color-CLC: The lines have the same color feature.

Orientation-CLC: The lines are parallel to each other or converge to a common vanishing point.

Spatially-CLC: The lines are in close proximity to each other.

57

Color-CLC

Color feature of lines: color pair (c1,c2)Color pair space: RGB (2563*2563) Too big!Dominant colors (20*20)

Finding the color pairs:One line → Several color pairs

Constructing Color-CLC: use clustering

58

Color-CLC

59

Orientation-CLC

The lines in an Orientation-CLC are parallel to each other in the 3D worldThe parallel lines of an object in a 2D image can be:

Parallel in 2DConverging to a vanishing point (perspective)

60

Orientation-CLC

61

Spatially-CLC

Vertical position clusteringHorizontal position clustering

62

Building Recognition by CLCTwo types of buildings → Two criteria

Inter-relationship criterionIntra-relationship criterion

63

Inter-relationship criterion(Nc1>Ti1 or Nc2>Ti1) and (Nc1+Nc2)>Ti2

Nc1 = number of intersecting lines in cluster 1

Nc2 = number of intersecting lines in cluster 2

64

Intra-relationship criterion

|So| > Tj1 or w(So) > Tj2

S0 = set of heavily overlapping lines in a cluster

65

Experimental EvaluationObject Recognition

97 well-patterned buildings (bp): 97/9744 not well-patterned buildings (bnp): 42/4416 not patterned non-buildings (nbnp): 15/16 (one false positive)25 patterned non-buildings (nbp): 0/25

CBIR

66

Experimental EvaluationWell-Patterned Buildings

67

Experimental EvaluationNon-Well-Patterned Buildings

68

Experimental EvaluationNon-Well-Patterned Non-Buildings

69

Experimental EvaluationWell-Patterned Non-Buildings (false positives)

70

Experimental Evaluation (CBIR)

Total Positive Classification

(#)

Total Negative

Classification(#)

False positive

(#)

False negative

(#)

Accuracy(%)

Arborgreens 0 47 0 0 100

89.6

87.5

91.7

Campusinfall 27 21 0 5

Cannonbeach 30 18 0 6

Yellowstone 4 44 4 0

71

Experimental Evaluation (CBIR)False positives from Yellowstone